Add context-delta prompt writing skill

[lastobelus]

What changes

• add a generic new-prompt skill for writing ready-to-paste prompts for fresh agents, subagents, restarted threads, and different environments
• make the recipient startup context the baseline and preserve the thread-specific context delta
• distinguish durable pointers, settled decisions, observations, attempts, hypotheses, and stale-able snapshots
• keep prompt writing separate from executing or delegating the prompt

Why

Agent-to-agent prompts frequently repeat global and repository guidance the recipient will already receive while dropping the decisions, failed experiments, discoveries, and exact continuation state unique to the originating thread. That wastes context and forces repeated investigation.

Boundary

Done when the skill validates and independent Markover-based trials show that generated prompts retain thread-only decisions, discoveries, attempt/results, and continuation state while using pointers for inherited or durable material. This PR excludes global installation and Markover application behavior.

Validation

• skill-creator quick_validate.py
• independent Markover forward trials
• repository-local checks

[lastobelus]

Work intent

phase: implementing
summary: "Develop and trial a generic new-prompt skill that transfers thread-specific context without recopying recipient startup guidance"
touch-points:
  - ".agents/skills/new-prompt/SKILL.md and agents/openai.yaml"
  - "isolated Markover-based forward trials and resulting focused revisions"
done-when:
  - "the skill passes skill-creator validation and repository-local checks"
  - "independent trials retain thread-only user decisions, costly discoveries, attempt/result evidence, and the exact continuation point"
  - "independent trials use purposeful pointers instead of paraphrasing inherited guidance or durable source contents"
  - "prompt-only requests return a ready-to-paste prompt without running or delegating it"
excludes:
  - "global skill installation or publication"
  - "Markover product or runtime behavior"
  - "a compatibility layer or generalized prompt-evaluation framework"
blocked-by: []
may-block: []
branch: "t3code/create-agent-prompt-skill"